Wuyxin/DISC
(ICML 2023) Discover and Cure: Concept-aware Mitigation of Spurious Correlation
This project helps machine learning practitioners build more reliable image classification models. It takes your existing image datasets and an initial classification model, and then identifies and corrects "spurious correlations"—situations where the model relies on irrelevant background details instead of the main subject. The output is a more accurate and trustworthy model that makes decisions based on the true content of images.
Use this if your image classification models are performing poorly on new, diverse images or if you suspect they are making decisions based on misleading background elements.
Not ideal if you are working with non-image data or if your primary concern is model interpretability without needing to retrain for improved robustness.
Stars
44
Forks
6
Language
Python
License
MIT
Category
Last pushed
Nov 17, 2025
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/Wuyxin/DISC"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
OPTML-Group/Unlearn-Saliency
[ICLR24 (Spotlight)] "SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in...
Shilin-LU/VINE
[ICLR 2025] "Robust Watermarking Using Generative Priors Against Image Editing: From...
WindVChen/DiffAttack
An unrestricted attack based on diffusion models that can achieve both good transferability and...
koninik/DiffusionPen
Official PyTorch Implementation of "DiffusionPen: Towards Controlling the Style of Handwritten...
bytedance/LatentUnfold
Implementation of paper: Flux Already Knows – Activating Subject-Driven Image Generation without Training